Research On Low-Quality Deepfake Face Detection Method

Rapid advances in computer graphics and deep learning techniques makes it possible for any user to fake realistic facial media content at will.The malicious dissemination of fake face images on Internet platforms has led to people no longer believe that seeing is believing.Therefore,deepfake detection has become one of the hot issues in the field of cyberspace security.Existing works use handcrafted features or fake patterns extracted from CNN networks to detect fake faces.However,the compression operation will mask some forged traces,making the traditional detection methods no longer effective.We find that it is easier to detect certain subtle properties in the spectrum of the input image than on its raw pixels,and existing face generation techniques do not enhance the temporal continuity of the video.Therefore,this dissertation solves the problem of low-quality deepfake face detection from two aspects:frequency feature extraction and spatio-temporal information modeling.1.We proposed a low-quality fake face image detection method based on frequency noise.Traditional methods lack the interaction between frequency and spatial domain.Therefore,this dissertation proposes three detection models that realize the interaction of those two domains,which can promote the feature extraction capability of frequency feature extractor and RGB feature extractor.Channel attention and spatial attention mechanisms are added to the model to further improve performance.In terms of experiments,the detection accuracy of this method on the four datasets of FaceForensics++has improved by an average of 1.6%compared with the traditional methods.2.We proposed a Multimodal low-quality fake face video detection method.The method is a novel two-stream structure.A spectral decomposition mechanism is proposed in the frequency stream to extract the low,medium and high frequency components of the input image.And a channel attention mechanism is introduced to adaptively learn the weights of different spectral components.The spatio-temporal stream mainly exploits separable 3D convolutions to extract spatio-temporal features between groups of frames.Finally,the two streams are fused to obtain the final detection result.In terms of experiments,the detection accuracy of this method on the four datasets of FaceForensics++has improved by an average of 1.8%compared with the traditional methods.